Solving the Inverse Problem of Analog Integrated Circuit Sizing with Diffusion Models

In this paper, the focus is given on using artificial neural networks (ANNs), particularly diffusion models, to automate the sizing of analog integrated circuits (ICs), given the constraints of its performance metrics. Traditionally, metaheuristics and optimization-based approaches have been explored to address this challenge, but each method has its drawbacks and yields inefficient results’ production. ANNs have been used in some works, but there is the common hurdle of small datasets to train the models in a supervised manner. In this work, we propose using denoising diffusion probabilistic models (DDPMs) to tackle the inverse problem of analog IC sizing. Several model architectures are trained to gradually learn to remove noise, meaning that after training, they can produce new data from random noise samples. We show that even a simple DDPM can sample sizing solutions in a small amount of time, which is an important stepping stone for future research.